mdmon.c

   1
   2 /*
   3  * md array manager.
   4  * When md arrays have user-space managed metadata, this is the program
   5  * that does the managing.
   6  *
   7  * Given one argument: the name of the array (e.g. /dev/md0) that is
   8  * the container.
   9  * We fork off a helper that runs high priority and mlocked.  It responds to
  10  * device failures and other events that might stop writeout, or that are
  11  * trivial to deal with.
  12  * The main thread then watches for new arrays being created in the container
  13  * and starts monitoring them too ... along with a few other tasks.
  14  *
  15  * The main thread communicates with the priority thread by writing over
  16  * a pipe.
  17  * Separate programs can communicate with the main thread via Unix-domain
  18  * socket.
  19  * The two threads share address space and open file table.
  20  *
  21  */
  22
  23 #ifndef _GNU_SOURCE
  24 #define _GNU_SOURCE
  25 #endif
  26
  27 #include        <unistd.h>
  28 #include        <stdlib.h>
  29 #include        <sys/stat.h>
  30 #include        <sys/socket.h>
  31 #include        <sys/un.h>
  32 #include        <sys/mman.h>
  33 #include        <stdio.h>
  34 #include        <errno.h>
  35 #include        <string.h>
  36 #include        <fcntl.h>
  37
  38 #include        <sched.h>
  39
  40 #include        "mdadm.h"
  41 #include        "mdmon.h"
  42
  43 struct active_array *array_list;
  44 struct active_array *discard_this;
  45 struct active_array *pending_discard;
  46
  47 int run_child(void *v)
  48 {
  49         struct supertype *c = v;
  50         do_monitor(c);
  51         return 0;
  52 }
  53
  54 int clone_monitor(struct supertype *container)
  55 {
  56         int pfd[2];
  57         static char stack[4096];
  58         int rv;
  59
  60         pipe(container->pipe);
  61
  62         rv = clone(run_child, stack+4096-64,
  63                    CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD,
  64                    container);
  65
  66         if (rv < 0)
  67                 return rv;
  68         return pfd[1];
  69 }
  70
  71 static struct superswitch *find_metadata_methods(char *vers)
  72 {
  73         if (strcmp(vers, "ddf") == 0)
  74                 return &super_ddf;
  75         return NULL;
  76 }
  77
  78
  79 static int make_pidfile(char *devname)
  80 {
  81         char path[100];
  82         char pid[10];
  83         int fd;
  84         sprintf(path, "/var/run/mdadm/%s.pid", devname);
  85
  86         fd = open(path, O_RDWR|O_CREAT|O_EXCL, 0600);
  87         if (fd < 0)
  88                 return -1;
  89         sprintf(pid, "%d\n", getpid());
  90         write(fd, pid, strlen(pid));
  91         close(fd);
  92         return 0;
  93 }
  94
  95 static int make_control_sock(char *devname)
  96 {
  97         char path[100];
  98         int sfd;
  99         long fl;
 100         struct sockaddr_un addr;
 101
 102         sprintf(path, "/var/run/mdadm/%s.sock", devname);
 103         unlink(path);
 104         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 105         if (sfd < 0)
 106                 return -1;
 107
 108         addr.sun_family = PF_LOCAL;
 109         strcpy(addr.sun_path, path);
 110         if (bind(sfd, &addr, sizeof(addr)) < 0) {
 111                 close(sfd);
 112                 return -1;
 113         }
 114         listen(sfd, 10);
 115         fl = fcntl(sfd, F_GETFL, 0);
 116         fl |= O_NONBLOCK;
 117         fcntl(sfd, F_SETFL, fl);
 118         return sfd;
 119 }
 120
 121 int main(int argc, char *argv[])
 122 {
 123         int mdfd;
 124         int pipefd;
 125         struct mdinfo *mdi, *di;
 126         struct supertype *container;
 127         if (argc != 2) {
 128                 fprintf(stderr, "Usage: md-manage /device/name/for/container\n");
 129                 exit(2);
 130         }
 131         mdfd = open(argv[1], O_RDWR);
 132         if (mdfd < 0) {
 133                 fprintf(stderr, "md-manage: %s: %s\n", argv[1],
 134                         strerror(errno));
 135                 exit(1);
 136         }
 137         if (md_get_version(mdfd) < 0) {
 138                 fprintf(stderr, "md-manage: %s: Not an md device\n",
 139                         argv[1]);
 140                 exit(1);
 141         }
 142
 143         /* hopefully it is a container - we'll check later */
 144
 145         container = malloc(sizeof(*container));
 146         container->devfd = mdfd;
 147         container->devnum = fd2devnum(mdfd);
 148         container->devname = devnum2devname(container->devnum);
 149
 150         /* If this fails, we hope it already exists */
 151         mkdir("/var/run/mdadm", 0600);
 152         /* pid file lives in /var/run/mdadm/mdXX.pid */
 153         if (make_pidfile(container->devname) < 0) {
 154                 fprintf(stderr, "md-manage: %s already managed\n",
 155                         container->devname);
 156                 exit(3);
 157         }
 158
 159         container->sock = make_control_sock(container->devname);
 160         if (container->sock < 0) {
 161                 fprintf(stderr, "mdmon: Cannot create socket in /var/run/mdadm\n");
 162                 exit(3);
 163         }
 164         container->arrays = NULL;
 165
 166         mdi = sysfs_read(mdfd, container->devnum,
 167                          GET_VERSION|GET_LEVEL|GET_DEVS);
 168
 169         if (!mdi) {
 170                 fprintf(stderr, "mdmon: failed to load sysfs info for %s\n",
 171                         container->devname);
 172                 exit(3);
 173         }
 174         if (mdi->array.level != UnSet) {
 175                 fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n",
 176                         argv[1]);
 177                 exit(3);
 178         }
 179         if (mdi->array.major_version != -1 ||
 180             mdi->array.minor_version != -2) {
 181                 fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n",
 182                         argv[1]);
 183                 exit(3);
 184         }
 185
 186         container->ss = find_metadata_methods(mdi->text_version);
 187         if (container->ss == NULL) {
 188                 fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n",
 189                         argv[1], mdi->text_version);
 190                 exit(3);
 191         }
 192
 193         container->devs = NULL;
 194         for (di = mdi->devs; di; di = di->next) {
 195                 struct mdinfo *cd = malloc(sizeof(*cd));
 196                 cd = di;
 197                 cd->next = container->devs;
 198                 container->devs = cd;
 199         }
 200         sysfs_free(mdi);
 201
 202
 203         if (container->ss->load_super(container, mdfd, argv[1])) {
 204                 fprintf(stderr, "mdmon: Cannot load metadata for %s\n",
 205                         argv[1]);
 206                 exit(3);
 207         }
 208
 209
 210         mlockall(MCL_FUTURE);
 211
 212         pipefd = clone_monitor(container);
 213         if (pipefd < 0) {
 214                 fprintf(stderr, "md-manage: failed to start monitor process: %s\n",
 215                         strerror(errno));
 216                 exit(2);
 217         }
 218
 219         do_manager(container);
 220
 221         exit(0);
 222 }